Meta's first frontier model bets on memory

Meta's Superintelligence Lab released Muse Wonder and a larger Muse Wonder Full on October 12, the company's first real attempt at a frontier model. On general benchmarks Meta says the models land around Fable 5.1 and GPT-6 Astra, and do better on harder tasks like coding. The number Meta wants you to notice is MemoryBench, a test of continual learning, where it says Muse Wonder beats Fable, Opus, and Astra. The pitch is "active memory": the model is built to hold on to what you tell it over weeks and pull the right pieces back later, instead of starting fresh each session.

That focus is not an accident. Both models will power Muse, Meta's personal agent, and an assistant that forgets your context every conversation is a weak agent. TechCrunch reports that Muse Wonder runs on Meta's Model API at $4 per million input tokens and $20 per million output tokens, roughly the same as GPT-6 Sol and Opus 5.5. Muse Wonder Full is held back for Muse Code subscribers and is not sold through the API. The models were trained at Prometheus, Meta's data center in central Ohio that came online earlier this year.

This is the first state-of-the-art model Meta has shipped since Alexandr Wang reorganized its AI work in August 2025. Until now that effort produced the lightweight Muse Spark and the Muse agent, but no model that competed at the top.

Why it matters

If you are choosing a model for an assistant that runs over days rather than single prompts, memory is the axis Meta is competing on, and the price matches the incumbents. Worth testing on your own long-running workflows before assuming the benchmark win holds up.